Skip to content

Three-Tier Grounding Model

Questions or comments?

Post them as a comment on the tracking issue -- requires a GitHub account with access to this repo.

Scope

A framework for how confidently the advisor can answer for a given country: Tier 1 — Codex-grounded (fully crawled, SME-reviewed, ingested; the advisor answers from curated, versioned Codex content), Tier 2 — configured-source fallback (approved source websites identified but not fully crawled/curated; the advisor pulls directly from those known-good sites and cites the web source), Tier 3 — open web (no sources configured at all; the advisor falls back to general web search with no curation behind it). This has existed informally; this Feature Set is what documents it as the model governing both how Curator prioritizes ingestion work and how confidently Advisor should present an answer.

Value

Gives both Curator and Advisor a single, shared model to reason about per-jurisdiction confidence, instead of each side implicitly assuming its own definition of "good enough" coverage.

Acceptance criteria

  • Every supported (or partially supported) jurisdiction has a documented, current tier assignment, derivable from its four-state curation lifecycle status (see ../four-state-curation-lifecycle/spec.md — State D implies Tier 1, State B/C implies Tier 2 if sources are configured, State A implies Tier 3).
  • The Codex Curator ops dashboard (../curator-ops-dashboard/spec.md) surfaces tier per country as part of its per-country breakdown.
  • Nova Policy Advisor's answer-time behavior consumes this tier assignment to decide how to cite a source (curated Codex citation vs. web-source citation vs. open-web, uncited) — coordinate directly with ../../nova-policy-advisor/conversational-policy-advisor/spec.md so the two Feature Sets don't each invent their own version of this model.
  • Which tier grounded a given advisor answer is not required to be surfaced to the end user by this Feature Set alone — that's explicitly an open item assigned to the Policy Advisor side (flagging a Tier 3 answer with lower confidence than Tier 1). This Feature only needs to make the tier determinable, not necessarily displayed.